Fleet bench

The fleet, at a glance

Suggested roles follow the measurements below. They are starting points for real-work tests.

┌──────────────────────┐     ╭─────────────────────╮
│ MINI / M4 / 16 GB    ├───▶ │ Background compute  │
└──────────────────────┘     ╰─────────────────────╯

┌──────────────────────┐     ╭─────────────────────╮
│ AIR 2 / M4 / 24 GB   ├───▶ │ Mobile + larger RAM │
└──────────────────────┘     ╰─────────────────────╯

┌──────────────────────┐     ╭─────────────────────╮
│ DESKTOP / 9900K      ├───▶ │ NVENC + gaming      │
└──────────────────────┘     ╰─────────────────────╯

┌──────────────────────┐     ╭─────────────────────╮
│ OMARCHY / M1 / 8 GB  ├───▶ │ Light Linux jobs    │
└──────────────────────┘     ╰─────────────────────╯

┌──────────────────────┐     ╭─────────────────────╮
│ YARD / 1 vCPU        ├───▶ │ Always-on services  │
└──────────────────────┘     ╰─────────────────────╯

╔════════════════════════════════════════════════╗
║ 385 accepted trials · 5 hosts · 2-hour window  ║
╚════════════════════════════════════════════════╝

┈┈ Follow-ups: games · real exports · concurrency

MINI / M4 / 16 GB
└──▶ Background compute

AIR 2 / M4 / 24 GB
└──▶ Mobile + larger RAM

DESKTOP / 9900K
└──▶ NVENC + gaming

OMARCHY / M1 / 8 GB
└──▶ Light Linux jobs

YARD / 1 vCPU
└──▶ Always-on services

╔════════════════════════════╗
║     385 accepted trials    ║
║ 5 hosts · a 2-hour window  ║
╚════════════════════════════╝

389 actual attempts across 420 planned cells. Two repaired runner failures added retries. Repeated brief work across two hours, not continuous maximum load. All benchmark cleanup was verified.

Spend where the benefit is clear

First preserve the music, then measure the job you want to improve. Cooler spending buys potential thermal or acoustic headroom; GPU and CPU spending address different throughput limits. Rehousing buys space. A PSU supports the selected hardware and is not the default first purchase.

╔══════════════════════════════╗
║ PRESERVE MUSIC FIRST         ║
╚══════════════════════════════╝
   ║ copies + restore checks
   ▼
┌──────────────────────────────┐
│ KEEP THE FLEET / $0 new gear │
└──────────────────────────────┘
   Mini / Air / Yard: reuse
   Desktop: 9900K + RTX 3060
   Omarchy: light Linux work

WHAT NEEDS TO IMPROVE?
├┈ Cooling or noise
│  Value air: PA120 Mini
│  $35–50 buying allowance
│  alt: NH-D12L $100–120
│  alt: Arctic 360 $110–160*
│  Choose ONE CPU cooler
├┈ Case airflow or fan noise
│  Keep current fans: $0
│  alt: P12 Pro PST 5-pack
│  $30–45 allowance if needed
├┈ GPU-limited games / compute
│  Used 4070 Ti SUPER 16 GB
│  Wait for ≤$700 target
├┈ CPU-limited work
│  Used 7950X+B650+64 GB
│  Wait for ≤$850 set target
├┈ New GPU needs another PSU
│  Confirm existing model
│  RM850x: $169.99 quote
└┈ Smaller footprint
   Used D41 Mesh: $50–80
   Space benefit; no speedup

┌──────────────────────────────┐
│ WINTER / ORIGINAL INTEL LAB  │
└──────────────────────────────┘
   RX 5600 XT + OpenCore
   Image / kexts / benchmarks

Dotted paths = optional choices.
*AIO + socket-kit allowance.
USD; targets are not offers.

All figures USD before tax/shipping unless specified. Allowances are spending plans; targets are buy thresholds; MSRP is a reference; earlier quotes require rechecking. No measured FPS-per-dollar or cooling-per-dollar comparison exists for these upgrades.

85°C was our stop point

All fans were already set to Full Speed. That change is complete. Desktop stopped in round 10’s four-worker CPU test, after nine accepted four-worker trials. Intel lists 100°C Tjunction for the 9900K; that specification is not a suggested temperature target. Intel specifications.

The defensible conclusion is narrower: we have not measured what this workload would do beyond our cutoff. A cooling upgrade is an option for more headroom or less noise. These results do not establish a required repair.

All measured results

Highest first within comparable conditions. Lower memory latency and network delay are better. No single score represents the whole machine.

One worker only. Macs were tested up to two workers and Desktop up to four; maximum all-core performance was not compared.

What stands out

• Mini and Air 2 are close in CPU work. Two-worker medians: 20.57 and 20.22 GFLOP/s. Both were repeatable under this limited load.
• Desktop leads this hardware-encode fixture. At 8M: 309.33 fps versus Mini 213.31 and Air 2 191.39. Outputs had different achieved bitrates and quality, so this is not an equal-quality speed contest.
• Desktop encoder timing varied. Its 8M range was 153.56–317.93 fps. A longer export will help separate short-process overhead from sustained behavior; the cause is not established.
• Desktop memory tells two stories. Native copy throughput was lower, while latency was better: 63.64 ns versus Mini 69.72 and Air 2 72.97. Neither finding justifies a manual multiplier increase.
• Small-host checks stayed deliberately small. Omarchy’s light runs did not repeat the earlier swap guard. Yard’s services stayed active. Neither result establishes higher safe concurrency.

Cooling: cost versus benefit

Cost/value changes the first cooler to price: Thermalright Peerless Assassin 120 Mini. The NH-D12L remains the premium compact-air choice. Neither has been tested on this desktop, so these are fit-and-budget recommendations, not a measured ranking of degrees per dollar.

The PA120 Mini has a 135 mm heatsink and supports LGA115X/AM5. The NH-D12L is 145 mm with its fan. The original O11 Dynamic permits 155 mm CPU coolers. A raised fan can consume clearance; check installed dimensions before ordering.

Arctic Liquid Freezer III Pro 360 is the liquid candidate that can span the current Intel platform and AM5. Budget $110–160 including a socket-kit allowance; local retail is unverified. Its thick radiator/fan assembly needs at least 63 mm clearance, plus tube, motherboard and cable room. Arctic’s LGA1200/115X kit supports the 9900K socket; confirm box revision and whether the kit must be purchased separately. Buy one CPU cooler, not air now plus liquid immediately afterward.

For liquid, the intended layout is top radiator exhaust with existing bottom/side fans supplying intake, subject to fit and actual fan placement. Keep the included radiator fans. Air cooling avoids pump maintenance; a 360 mm radiator offers more heat-rejection area for sustained work, with added installation and pump complexity. Neither guarantees silence or a specific temperature drop.

Case-fan choices: Arctic P12 Pro PST for value; Noctua NF-A12x25 G2 PWM for premium refinement. The Noctua launch MSRP was $34.90 each, not a verified checkout price. Existing fan RPM readings alone cannot establish airflow or acoustic quality. Test the chosen CPU cooler first, then spend on case fans only if a problem remains.

All fans are already Full Speed. The 85°C campaign cutoff was our policy, not evidence of a defective cooler or PSU. More FPS is possible only if cooling removes an actual thermal constraint; none has been demonstrated here. Compare identical load, room conditions, power, temperatures, clocks and noise after one change.

Next steps, in order

1. Benchmark an actual game

   Choose an installed game with a repeatable benchmark. Fix resolution, preset, upscaling, VSync and caps. Run one warm-up and three measured passes. Capture average FPS, 1% lows or frame times, temperatures, power, effective clocks and thermal-throttle flags.

   Decision: GPU limit, CPU limit, or acceptable performance? Preserve the existing guards unless the next protocol explicitly changes them. A guard stop remains an incomplete measurement, not a hardware diagnosis.

2. Compare one real export across three machines

   Use the same source with motion, detail and representative effects. Run three repetitions on Mini, Air 2 and Desktop at the same intended output quality. Review the outputs visually, then measure source transfer + queue + encode + result transfer separately.

   Decision: where does acceptable work finish fastest?

3. Measure transfers and useful worker counts

   Measure an Air 2 file transfer before changing its relay path. On Mini and Air 2, compare two versus four workers with the same job and pressure guards. Keep Yard’s service active and check response time with light background work. Keep Omarchy’s foreground apps and swap activity visible.

   Decision: useful concurrency without degrading interactive work or services. These follow-ups are proposed, not running.

4. Choose cooling around noise and sustained work

   If Desktop is too loud or the desired workload needs more headroom, choose a fitting cooler, record a baseline, install with fresh paste according to its manual, and repeat the same workload at comparable ambient conditions and package power. Measure actual RPM if available.

   Optional: compare normal versus full-speed fans under matched conditions. Repeating the completed BIOS fan change is not a prerequisite.

5. Keep full-memory qualification deferred

   The boot-time RAM test stays deferred as requested. Keep the selected XMP profile and memory multiplier Auto. We have not certified all 48 GiB; earlier Windows updates confound an XMP-only before/after claim.

Evidence and limits

This includes every metric family from the completed two-hour campaign, the earlier fan comparison, and the revised next steps. Tables use accepted trials; stopped runs do not become completed throughput scores. Earlier exploratory runs are context, not pooled into final medians.

No game FPS, wall-power efficiency, full-memory certification or continuous maximum-load qualification was measured. Background activity, operating systems, memory-test sizes and encoder backends limit cross-machine comparisons.

• Download aggregate result matrix (JSON) — counts, medians, ranges and per-trial values; no addresses or raw logs.
• Plain-text fleet diagram
• Intel i9-9900K specifications
• Original Lian Li O11 Dynamic case
• Noctua NH-D12L dimensions and sockets
• Thermalright Peerless Assassin 120 Mini
• Cooler Master G100M manufacturer datasheet
• Tailscale connection paths